xref: /netbsd-src/sys/kern/vfs_bio.c (revision 404fbe5fb94ca1e054339640cabb2801ce52dd30)
1 /*	$NetBSD: vfs_bio.c,v 1.215 2008/12/07 20:58:46 pooka Exp $	*/
2 
3 /*-
4  * Copyright (c) 2007, 2008 The NetBSD Foundation, Inc.
5  * All rights reserved.
6  *
7  * This code is derived from software contributed to The NetBSD Foundation
8  * by Andrew Doran.
9  * This code is derived from software contributed to The NetBSD Foundation
10  * by Wasabi Systems, Inc.
11  *
12  * Redistribution and use in source and binary forms, with or without
13  * modification, are permitted provided that the following conditions
14  * are met:
15  * 1. Redistributions of source code must retain the above copyright
16  *    notice, this list of conditions and the following disclaimer.
17  * 2. Redistributions in binary form must reproduce the above copyright
18  *    notice, this list of conditions and the following disclaimer in the
19  *    documentation and/or other materials provided with the distribution.
20  *
21  * THIS SOFTWARE IS PROVIDED BY THE NETBSD FOUNDATION, INC. AND CONTRIBUTORS
22  * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED
23  * TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
24  * PURPOSE ARE DISCLAIMED.  IN NO EVENT SHALL THE FOUNDATION OR CONTRIBUTORS
25  * BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
26  * CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
27  * SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
28  * INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
29  * CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
30  * ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
31  * POSSIBILITY OF SUCH DAMAGE.
32  */
33 
34 /*-
35  * Copyright (c) 1982, 1986, 1989, 1993
36  *	The Regents of the University of California.  All rights reserved.
37  * (c) UNIX System Laboratories, Inc.
38  * All or some portions of this file are derived from material licensed
39  * to the University of California by American Telephone and Telegraph
40  * Co. or Unix System Laboratories, Inc. and are reproduced herein with
41  * the permission of UNIX System Laboratories, Inc.
42  *
43  * Redistribution and use in source and binary forms, with or without
44  * modification, are permitted provided that the following conditions
45  * are met:
46  * 1. Redistributions of source code must retain the above copyright
47  *    notice, this list of conditions and the following disclaimer.
48  * 2. Redistributions in binary form must reproduce the above copyright
49  *    notice, this list of conditions and the following disclaimer in the
50  *    documentation and/or other materials provided with the distribution.
51  * 3. Neither the name of the University nor the names of its contributors
52  *    may be used to endorse or promote products derived from this software
53  *    without specific prior written permission.
54  *
55  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
56  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
57  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
58  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
59  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
60  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
61  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
62  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
63  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
64  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
65  * SUCH DAMAGE.
66  *
67  *	@(#)vfs_bio.c	8.6 (Berkeley) 1/11/94
68  */
69 
70 /*-
71  * Copyright (c) 1994 Christopher G. Demetriou
72  *
73  * Redistribution and use in source and binary forms, with or without
74  * modification, are permitted provided that the following conditions
75  * are met:
76  * 1. Redistributions of source code must retain the above copyright
77  *    notice, this list of conditions and the following disclaimer.
78  * 2. Redistributions in binary form must reproduce the above copyright
79  *    notice, this list of conditions and the following disclaimer in the
80  *    documentation and/or other materials provided with the distribution.
81  * 3. All advertising materials mentioning features or use of this software
82  *    must display the following acknowledgement:
83  *	This product includes software developed by the University of
84  *	California, Berkeley and its contributors.
85  * 4. Neither the name of the University nor the names of its contributors
86  *    may be used to endorse or promote products derived from this software
87  *    without specific prior written permission.
88  *
89  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
90  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
91  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
92  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
93  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
94  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
95  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
96  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
97  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
98  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
99  * SUCH DAMAGE.
100  *
101  *	@(#)vfs_bio.c	8.6 (Berkeley) 1/11/94
102  */
103 
104 /*
105  * Some references:
106  *	Bach: The Design of the UNIX Operating System (Prentice Hall, 1986)
107  *	Leffler, et al.: The Design and Implementation of the 4.3BSD
108  *		UNIX Operating System (Addison Welley, 1989)
109  */
110 
111 #include <sys/cdefs.h>
112 __KERNEL_RCSID(0, "$NetBSD: vfs_bio.c,v 1.215 2008/12/07 20:58:46 pooka Exp $");
113 
114 #include "fs_ffs.h"
115 #include "opt_bufcache.h"
116 
117 #include <sys/param.h>
118 #include <sys/systm.h>
119 #include <sys/kernel.h>
120 #include <sys/proc.h>
121 #include <sys/buf.h>
122 #include <sys/vnode.h>
123 #include <sys/mount.h>
124 #include <sys/resourcevar.h>
125 #include <sys/sysctl.h>
126 #include <sys/conf.h>
127 #include <sys/kauth.h>
128 #include <sys/fstrans.h>
129 #include <sys/intr.h>
130 #include <sys/cpu.h>
131 #include <sys/wapbl.h>
132 
133 #include <uvm/uvm.h>
134 
135 #include <miscfs/specfs/specdev.h>
136 
137 #ifndef	BUFPAGES
138 # define BUFPAGES 0
139 #endif
140 
141 #ifdef BUFCACHE
142 # if (BUFCACHE < 5) || (BUFCACHE > 95)
143 #  error BUFCACHE is not between 5 and 95
144 # endif
145 #else
146 # define BUFCACHE 15
147 #endif
148 
149 u_int	nbuf;			/* XXX - for softdep_lockedbufs */
150 u_int	bufpages = BUFPAGES;	/* optional hardwired count */
151 u_int	bufcache = BUFCACHE;	/* max % of RAM to use for buffer cache */
152 
153 /* Function prototypes */
154 struct bqueue;
155 
156 static void buf_setwm(void);
157 static int buf_trim(void);
158 static void *bufpool_page_alloc(struct pool *, int);
159 static void bufpool_page_free(struct pool *, void *);
160 static buf_t *bio_doread(struct vnode *, daddr_t, int,
161     kauth_cred_t, int);
162 static buf_t *getnewbuf(int, int, int);
163 static int buf_lotsfree(void);
164 static int buf_canrelease(void);
165 static u_long buf_mempoolidx(u_long);
166 static u_long buf_roundsize(u_long);
167 static void *buf_malloc(size_t);
168 static void buf_mrelease(void *, size_t);
169 static void binsheadfree(buf_t *, struct bqueue *);
170 static void binstailfree(buf_t *, struct bqueue *);
171 int count_lock_queue(void); /* XXX */
172 #ifdef DEBUG
173 static int checkfreelist(buf_t *, struct bqueue *, int);
174 #endif
175 static void biointr(void *);
176 static void biodone2(buf_t *);
177 static void bref(buf_t *);
178 static void brele(buf_t *);
179 static void sysctl_kern_buf_setup(void);
180 static void sysctl_vm_buf_setup(void);
181 
182 /*
183  * Definitions for the buffer hash lists.
184  */
185 #define	BUFHASH(dvp, lbn)	\
186 	(&bufhashtbl[(((long)(dvp) >> 8) + (int)(lbn)) & bufhash])
187 LIST_HEAD(bufhashhdr, buf) *bufhashtbl, invalhash;
188 u_long	bufhash;
189 struct bqueue bufqueues[BQUEUES];
190 const struct bio_ops *bioopsp;	/* I/O operation notification */
191 
192 static kcondvar_t needbuffer_cv;
193 
194 /*
195  * Buffer queue lock.
196  */
197 kmutex_t bufcache_lock;
198 kmutex_t buffer_lock;
199 
200 /* Software ISR for completed transfers. */
201 static void *biodone_sih;
202 
203 /* Buffer pool for I/O buffers. */
204 static pool_cache_t buf_cache;
205 static pool_cache_t bufio_cache;
206 
207 /* XXX - somewhat gross.. */
208 #if MAXBSIZE == 0x2000
209 #define NMEMPOOLS 5
210 #elif MAXBSIZE == 0x4000
211 #define NMEMPOOLS 6
212 #elif MAXBSIZE == 0x8000
213 #define NMEMPOOLS 7
214 #else
215 #define NMEMPOOLS 8
216 #endif
217 
218 #define MEMPOOL_INDEX_OFFSET 9	/* smallest pool is 512 bytes */
219 #if (1 << (NMEMPOOLS + MEMPOOL_INDEX_OFFSET - 1)) != MAXBSIZE
220 #error update vfs_bio buffer memory parameters
221 #endif
222 
223 /* Buffer memory pools */
224 static struct pool bmempools[NMEMPOOLS];
225 
226 static struct vm_map *buf_map;
227 
228 /*
229  * Buffer memory pool allocator.
230  */
231 static void *
232 bufpool_page_alloc(struct pool *pp, int flags)
233 {
234 
235 	return (void *)uvm_km_alloc(buf_map,
236 	    MAXBSIZE, MAXBSIZE,
237 	    ((flags & PR_WAITOK) ? 0 : UVM_KMF_NOWAIT | UVM_KMF_TRYLOCK)
238 	    | UVM_KMF_WIRED);
239 }
240 
241 static void
242 bufpool_page_free(struct pool *pp, void *v)
243 {
244 
245 	uvm_km_free(buf_map, (vaddr_t)v, MAXBSIZE, UVM_KMF_WIRED);
246 }
247 
248 static struct pool_allocator bufmempool_allocator = {
249 	.pa_alloc = bufpool_page_alloc,
250 	.pa_free = bufpool_page_free,
251 	.pa_pagesz = MAXBSIZE,
252 };
253 
254 /* Buffer memory management variables */
255 u_long bufmem_valimit;
256 u_long bufmem_hiwater;
257 u_long bufmem_lowater;
258 u_long bufmem;
259 
260 /*
261  * MD code can call this to set a hard limit on the amount
262  * of virtual memory used by the buffer cache.
263  */
264 int
265 buf_setvalimit(vsize_t sz)
266 {
267 
268 	/* We need to accommodate at least NMEMPOOLS of MAXBSIZE each */
269 	if (sz < NMEMPOOLS * MAXBSIZE)
270 		return EINVAL;
271 
272 	bufmem_valimit = sz;
273 	return 0;
274 }
275 
276 static void
277 buf_setwm(void)
278 {
279 
280 	bufmem_hiwater = buf_memcalc();
281 	/* lowater is approx. 2% of memory (with bufcache = 15) */
282 #define	BUFMEM_WMSHIFT	3
283 #define	BUFMEM_HIWMMIN	(64 * 1024 << BUFMEM_WMSHIFT)
284 	if (bufmem_hiwater < BUFMEM_HIWMMIN)
285 		/* Ensure a reasonable minimum value */
286 		bufmem_hiwater = BUFMEM_HIWMMIN;
287 	bufmem_lowater = bufmem_hiwater >> BUFMEM_WMSHIFT;
288 }
289 
290 #ifdef DEBUG
291 int debug_verify_freelist = 0;
292 static int
293 checkfreelist(buf_t *bp, struct bqueue *dp, int ison)
294 {
295 	buf_t *b;
296 
297 	if (!debug_verify_freelist)
298 		return 1;
299 
300 	TAILQ_FOREACH(b, &dp->bq_queue, b_freelist) {
301 		if (b == bp)
302 			return ison ? 1 : 0;
303 	}
304 
305 	return ison ? 0 : 1;
306 }
307 #endif
308 
309 /*
310  * Insq/Remq for the buffer hash lists.
311  * Call with buffer queue locked.
312  */
313 static void
314 binsheadfree(buf_t *bp, struct bqueue *dp)
315 {
316 
317 	KASSERT(mutex_owned(&bufcache_lock));
318 	KASSERT(bp->b_freelistindex == -1);
319 	TAILQ_INSERT_HEAD(&dp->bq_queue, bp, b_freelist);
320 	dp->bq_bytes += bp->b_bufsize;
321 	bp->b_freelistindex = dp - bufqueues;
322 }
323 
324 static void
325 binstailfree(buf_t *bp, struct bqueue *dp)
326 {
327 
328 	KASSERT(mutex_owned(&bufcache_lock));
329 	KASSERT(bp->b_freelistindex == -1);
330 	TAILQ_INSERT_TAIL(&dp->bq_queue, bp, b_freelist);
331 	dp->bq_bytes += bp->b_bufsize;
332 	bp->b_freelistindex = dp - bufqueues;
333 }
334 
335 void
336 bremfree(buf_t *bp)
337 {
338 	struct bqueue *dp;
339 	int bqidx = bp->b_freelistindex;
340 
341 	KASSERT(mutex_owned(&bufcache_lock));
342 
343 	KASSERT(bqidx != -1);
344 	dp = &bufqueues[bqidx];
345 	KDASSERT(checkfreelist(bp, dp, 1));
346 	KASSERT(dp->bq_bytes >= bp->b_bufsize);
347 	TAILQ_REMOVE(&dp->bq_queue, bp, b_freelist);
348 	dp->bq_bytes -= bp->b_bufsize;
349 
350 	/* For the sysctl helper. */
351 	if (bp == dp->bq_marker)
352 		dp->bq_marker = NULL;
353 
354 #if defined(DIAGNOSTIC)
355 	bp->b_freelistindex = -1;
356 #endif /* defined(DIAGNOSTIC) */
357 }
358 
359 /*
360  * Add a reference to an buffer structure that came from buf_cache.
361  */
362 static inline void
363 bref(buf_t *bp)
364 {
365 
366 	KASSERT(mutex_owned(&bufcache_lock));
367 	KASSERT(bp->b_refcnt > 0);
368 
369 	bp->b_refcnt++;
370 }
371 
372 /*
373  * Free an unused buffer structure that came from buf_cache.
374  */
375 static inline void
376 brele(buf_t *bp)
377 {
378 
379 	KASSERT(mutex_owned(&bufcache_lock));
380 	KASSERT(bp->b_refcnt > 0);
381 
382 	if (bp->b_refcnt-- == 1) {
383 		buf_destroy(bp);
384 #ifdef DEBUG
385 		memset((char *)bp, 0, sizeof(*bp));
386 #endif
387 		pool_cache_put(buf_cache, bp);
388 	}
389 }
390 
391 /*
392  * note that for some ports this is used by pmap bootstrap code to
393  * determine kva size.
394  */
395 u_long
396 buf_memcalc(void)
397 {
398 	u_long n;
399 
400 	/*
401 	 * Determine the upper bound of memory to use for buffers.
402 	 *
403 	 *	- If bufpages is specified, use that as the number
404 	 *	  pages.
405 	 *
406 	 *	- Otherwise, use bufcache as the percentage of
407 	 *	  physical memory.
408 	 */
409 	if (bufpages != 0) {
410 		n = bufpages;
411 	} else {
412 		if (bufcache < 5) {
413 			printf("forcing bufcache %d -> 5", bufcache);
414 			bufcache = 5;
415 		}
416 		if (bufcache > 95) {
417 			printf("forcing bufcache %d -> 95", bufcache);
418 			bufcache = 95;
419 		}
420 		n = calc_cache_size(buf_map, bufcache,
421 		    (buf_map != kernel_map) ? 100 : BUFCACHE_VA_MAXPCT)
422 		    / PAGE_SIZE;
423 	}
424 
425 	n <<= PAGE_SHIFT;
426 	if (bufmem_valimit != 0 && n > bufmem_valimit)
427 		n = bufmem_valimit;
428 
429 	return (n);
430 }
431 
432 /*
433  * Initialize buffers and hash links for buffers.
434  */
435 void
436 bufinit(void)
437 {
438 	struct bqueue *dp;
439 	int use_std;
440 	u_int i;
441 
442 	mutex_init(&bufcache_lock, MUTEX_DEFAULT, IPL_NONE);
443 	mutex_init(&buffer_lock, MUTEX_DEFAULT, IPL_NONE);
444 	cv_init(&needbuffer_cv, "needbuf");
445 
446 	if (bufmem_valimit != 0) {
447 		vaddr_t minaddr = 0, maxaddr;
448 		buf_map = uvm_km_suballoc(kernel_map, &minaddr, &maxaddr,
449 					  bufmem_valimit, 0, false, 0);
450 		if (buf_map == NULL)
451 			panic("bufinit: cannot allocate submap");
452 	} else
453 		buf_map = kernel_map;
454 
455 	/*
456 	 * Initialize buffer cache memory parameters.
457 	 */
458 	bufmem = 0;
459 	buf_setwm();
460 
461 	/* On "small" machines use small pool page sizes where possible */
462 	use_std = (physmem < atop(16*1024*1024));
463 
464 	/*
465 	 * Also use them on systems that can map the pool pages using
466 	 * a direct-mapped segment.
467 	 */
468 #ifdef PMAP_MAP_POOLPAGE
469 	use_std = 1;
470 #endif
471 
472 	buf_cache = pool_cache_init(sizeof(buf_t), 0, 0, 0,
473 	    "bufpl", NULL, IPL_SOFTBIO, NULL, NULL, NULL);
474 	bufio_cache = pool_cache_init(sizeof(buf_t), 0, 0, 0,
475 	    "biopl", NULL, IPL_BIO, NULL, NULL, NULL);
476 
477 	bufmempool_allocator.pa_backingmap = buf_map;
478 	for (i = 0; i < NMEMPOOLS; i++) {
479 		struct pool_allocator *pa;
480 		struct pool *pp = &bmempools[i];
481 		u_int size = 1 << (i + MEMPOOL_INDEX_OFFSET);
482 		char *name = kmem_alloc(8, KM_SLEEP);
483 		if (__predict_true(size >= 1024))
484 			(void)snprintf(name, 8, "buf%dk", size / 1024);
485 		else
486 			(void)snprintf(name, 8, "buf%db", size);
487 		pa = (size <= PAGE_SIZE && use_std)
488 			? &pool_allocator_nointr
489 			: &bufmempool_allocator;
490 		pool_init(pp, size, 0, 0, 0, name, pa, IPL_NONE);
491 		pool_setlowat(pp, 1);
492 		pool_sethiwat(pp, 1);
493 	}
494 
495 	/* Initialize the buffer queues */
496 	for (dp = bufqueues; dp < &bufqueues[BQUEUES]; dp++) {
497 		TAILQ_INIT(&dp->bq_queue);
498 		dp->bq_bytes = 0;
499 	}
500 
501 	/*
502 	 * Estimate hash table size based on the amount of memory we
503 	 * intend to use for the buffer cache. The average buffer
504 	 * size is dependent on our clients (i.e. filesystems).
505 	 *
506 	 * For now, use an empirical 3K per buffer.
507 	 */
508 	nbuf = (bufmem_hiwater / 1024) / 3;
509 	bufhashtbl = hashinit(nbuf, HASH_LIST, true, &bufhash);
510 
511 	sysctl_kern_buf_setup();
512 	sysctl_vm_buf_setup();
513 }
514 
515 void
516 bufinit2(void)
517 {
518 
519 	biodone_sih = softint_establish(SOFTINT_BIO | SOFTINT_MPSAFE, biointr,
520 	    NULL);
521 	if (biodone_sih == NULL)
522 		panic("bufinit2: can't establish soft interrupt");
523 }
524 
525 static int
526 buf_lotsfree(void)
527 {
528 	int try, thresh;
529 
530 	/* Always allocate if less than the low water mark. */
531 	if (bufmem < bufmem_lowater)
532 		return 1;
533 
534 	/* Never allocate if greater than the high water mark. */
535 	if (bufmem > bufmem_hiwater)
536 		return 0;
537 
538 	/* If there's anything on the AGE list, it should be eaten. */
539 	if (TAILQ_FIRST(&bufqueues[BQ_AGE].bq_queue) != NULL)
540 		return 0;
541 
542 	/*
543 	 * The probabily of getting a new allocation is inversely
544 	 * proportional to the current size of the cache, using
545 	 * a granularity of 16 steps.
546 	 */
547 	try = random() & 0x0000000fL;
548 
549 	/* Don't use "16 * bufmem" here to avoid a 32-bit overflow. */
550 	thresh = (bufmem - bufmem_lowater) /
551 	    ((bufmem_hiwater - bufmem_lowater) / 16);
552 
553 	if (try >= thresh)
554 		return 1;
555 
556 	/* Otherwise don't allocate. */
557 	return 0;
558 }
559 
560 /*
561  * Return estimate of bytes we think need to be
562  * released to help resolve low memory conditions.
563  *
564  * => called with bufcache_lock held.
565  */
566 static int
567 buf_canrelease(void)
568 {
569 	int pagedemand, ninvalid = 0;
570 
571 	KASSERT(mutex_owned(&bufcache_lock));
572 
573 	if (bufmem < bufmem_lowater)
574 		return 0;
575 
576 	if (bufmem > bufmem_hiwater)
577 		return bufmem - bufmem_hiwater;
578 
579 	ninvalid += bufqueues[BQ_AGE].bq_bytes;
580 
581 	pagedemand = uvmexp.freetarg - uvmexp.free;
582 	if (pagedemand < 0)
583 		return ninvalid;
584 	return MAX(ninvalid, MIN(2 * MAXBSIZE,
585 	    MIN((bufmem - bufmem_lowater) / 16, pagedemand * PAGE_SIZE)));
586 }
587 
588 /*
589  * Buffer memory allocation helper functions
590  */
591 static u_long
592 buf_mempoolidx(u_long size)
593 {
594 	u_int n = 0;
595 
596 	size -= 1;
597 	size >>= MEMPOOL_INDEX_OFFSET;
598 	while (size) {
599 		size >>= 1;
600 		n += 1;
601 	}
602 	if (n >= NMEMPOOLS)
603 		panic("buf mem pool index %d", n);
604 	return n;
605 }
606 
607 static u_long
608 buf_roundsize(u_long size)
609 {
610 	/* Round up to nearest power of 2 */
611 	return (1 << (buf_mempoolidx(size) + MEMPOOL_INDEX_OFFSET));
612 }
613 
614 static void *
615 buf_malloc(size_t size)
616 {
617 	u_int n = buf_mempoolidx(size);
618 	void *addr;
619 
620 	while (1) {
621 		addr = pool_get(&bmempools[n], PR_NOWAIT);
622 		if (addr != NULL)
623 			break;
624 
625 		/* No memory, see if we can free some. If so, try again */
626 		mutex_enter(&bufcache_lock);
627 		if (buf_drain(1) > 0) {
628 			mutex_exit(&bufcache_lock);
629 			continue;
630 		}
631 
632 		if (curlwp == uvm.pagedaemon_lwp) {
633 			mutex_exit(&bufcache_lock);
634 			return NULL;
635 		}
636 
637 		/* Wait for buffers to arrive on the LRU queue */
638 		cv_timedwait(&needbuffer_cv, &bufcache_lock, hz / 4);
639 		mutex_exit(&bufcache_lock);
640 	}
641 
642 	return addr;
643 }
644 
645 static void
646 buf_mrelease(void *addr, size_t size)
647 {
648 
649 	pool_put(&bmempools[buf_mempoolidx(size)], addr);
650 }
651 
652 /*
653  * bread()/breadn() helper.
654  */
655 static buf_t *
656 bio_doread(struct vnode *vp, daddr_t blkno, int size, kauth_cred_t cred,
657     int async)
658 {
659 	buf_t *bp;
660 	struct mount *mp;
661 
662 	bp = getblk(vp, blkno, size, 0, 0);
663 
664 #ifdef DIAGNOSTIC
665 	if (bp == NULL) {
666 		panic("bio_doread: no such buf");
667 	}
668 #endif
669 
670 	/*
671 	 * If buffer does not have data valid, start a read.
672 	 * Note that if buffer is BC_INVAL, getblk() won't return it.
673 	 * Therefore, it's valid if its I/O has completed or been delayed.
674 	 */
675 	if (!ISSET(bp->b_oflags, (BO_DONE | BO_DELWRI))) {
676 		/* Start I/O for the buffer. */
677 		SET(bp->b_flags, B_READ | async);
678 		if (async)
679 			BIO_SETPRIO(bp, BPRIO_TIMELIMITED);
680 		else
681 			BIO_SETPRIO(bp, BPRIO_TIMECRITICAL);
682 		VOP_STRATEGY(vp, bp);
683 
684 		/* Pay for the read. */
685 		curlwp->l_ru.ru_inblock++;
686 	} else if (async)
687 		brelse(bp, 0);
688 
689 	if (vp->v_type == VBLK)
690 		mp = vp->v_specmountpoint;
691 	else
692 		mp = vp->v_mount;
693 
694 	/*
695 	 * Collect statistics on synchronous and asynchronous reads.
696 	 * Reads from block devices are charged to their associated
697 	 * filesystem (if any).
698 	 */
699 	if (mp != NULL) {
700 		if (async == 0)
701 			mp->mnt_stat.f_syncreads++;
702 		else
703 			mp->mnt_stat.f_asyncreads++;
704 	}
705 
706 	return (bp);
707 }
708 
709 /*
710  * Read a disk block.
711  * This algorithm described in Bach (p.54).
712  */
713 int
714 bread(struct vnode *vp, daddr_t blkno, int size, kauth_cred_t cred,
715     int flags, buf_t **bpp)
716 {
717 	buf_t *bp;
718 	int error;
719 
720 	/* Get buffer for block. */
721 	bp = *bpp = bio_doread(vp, blkno, size, cred, 0);
722 
723 	/* Wait for the read to complete, and return result. */
724 	error = biowait(bp);
725 	if (error == 0 && (flags & B_MODIFY) != 0)	/* XXXX before the next code block or after? */
726 		error = fscow_run(bp, true);
727 
728 	return error;
729 }
730 
731 /*
732  * Read-ahead multiple disk blocks. The first is sync, the rest async.
733  * Trivial modification to the breada algorithm presented in Bach (p.55).
734  */
735 int
736 breadn(struct vnode *vp, daddr_t blkno, int size, daddr_t *rablks,
737     int *rasizes, int nrablks, kauth_cred_t cred, int flags, buf_t **bpp)
738 {
739 	buf_t *bp;
740 	int error, i;
741 
742 	bp = *bpp = bio_doread(vp, blkno, size, cred, 0);
743 
744 	/*
745 	 * For each of the read-ahead blocks, start a read, if necessary.
746 	 */
747 	mutex_enter(&bufcache_lock);
748 	for (i = 0; i < nrablks; i++) {
749 		/* If it's in the cache, just go on to next one. */
750 		if (incore(vp, rablks[i]))
751 			continue;
752 
753 		/* Get a buffer for the read-ahead block */
754 		mutex_exit(&bufcache_lock);
755 		(void) bio_doread(vp, rablks[i], rasizes[i], cred, B_ASYNC);
756 		mutex_enter(&bufcache_lock);
757 	}
758 	mutex_exit(&bufcache_lock);
759 
760 	/* Otherwise, we had to start a read for it; wait until it's valid. */
761 	error = biowait(bp);
762 	if (error == 0 && (flags & B_MODIFY) != 0)
763 		error = fscow_run(bp, true);
764 	return error;
765 }
766 
767 /*
768  * Read with single-block read-ahead.  Defined in Bach (p.55), but
769  * implemented as a call to breadn().
770  * XXX for compatibility with old file systems.
771  */
772 int
773 breada(struct vnode *vp, daddr_t blkno, int size, daddr_t rablkno,
774     int rabsize, kauth_cred_t cred, int flags, buf_t **bpp)
775 {
776 
777 	return (breadn(vp, blkno, size, &rablkno, &rabsize, 1,
778 	    cred, flags, bpp));
779 }
780 
781 /*
782  * Block write.  Described in Bach (p.56)
783  */
784 int
785 bwrite(buf_t *bp)
786 {
787 	int rv, sync, wasdelayed;
788 	struct vnode *vp;
789 	struct mount *mp;
790 
791 	KASSERT(ISSET(bp->b_cflags, BC_BUSY));
792 	KASSERT(!cv_has_waiters(&bp->b_done));
793 
794 	vp = bp->b_vp;
795 	if (vp != NULL) {
796 		KASSERT(bp->b_objlock == &vp->v_interlock);
797 		if (vp->v_type == VBLK)
798 			mp = vp->v_specmountpoint;
799 		else
800 			mp = vp->v_mount;
801 	} else {
802 		mp = NULL;
803 	}
804 
805 	if (mp && mp->mnt_wapbl) {
806 		if (bp->b_iodone != mp->mnt_wapbl_op->wo_wapbl_biodone) {
807 			bdwrite(bp);
808 			return 0;
809 		}
810 	}
811 
812 	/*
813 	 * Remember buffer type, to switch on it later.  If the write was
814 	 * synchronous, but the file system was mounted with MNT_ASYNC,
815 	 * convert it to a delayed write.
816 	 * XXX note that this relies on delayed tape writes being converted
817 	 * to async, not sync writes (which is safe, but ugly).
818 	 */
819 	sync = !ISSET(bp->b_flags, B_ASYNC);
820 	if (sync && mp != NULL && ISSET(mp->mnt_flag, MNT_ASYNC)) {
821 		bdwrite(bp);
822 		return (0);
823 	}
824 
825 	/*
826 	 * Collect statistics on synchronous and asynchronous writes.
827 	 * Writes to block devices are charged to their associated
828 	 * filesystem (if any).
829 	 */
830 	if (mp != NULL) {
831 		if (sync)
832 			mp->mnt_stat.f_syncwrites++;
833 		else
834 			mp->mnt_stat.f_asyncwrites++;
835 	}
836 
837 	/*
838 	 * Pay for the I/O operation and make sure the buf is on the correct
839 	 * vnode queue.
840 	 */
841 	bp->b_error = 0;
842 	wasdelayed = ISSET(bp->b_oflags, BO_DELWRI);
843 	CLR(bp->b_flags, B_READ);
844 	if (wasdelayed) {
845 		mutex_enter(&bufcache_lock);
846 		mutex_enter(bp->b_objlock);
847 		CLR(bp->b_oflags, BO_DONE | BO_DELWRI);
848 		reassignbuf(bp, bp->b_vp);
849 		mutex_exit(&bufcache_lock);
850 	} else {
851 		curlwp->l_ru.ru_oublock++;
852 		mutex_enter(bp->b_objlock);
853 		CLR(bp->b_oflags, BO_DONE | BO_DELWRI);
854 	}
855 	if (vp != NULL)
856 		vp->v_numoutput++;
857 	mutex_exit(bp->b_objlock);
858 
859 	/* Initiate disk write. */
860 	if (sync)
861 		BIO_SETPRIO(bp, BPRIO_TIMECRITICAL);
862 	else
863 		BIO_SETPRIO(bp, BPRIO_TIMELIMITED);
864 
865 	VOP_STRATEGY(vp, bp);
866 
867 	if (sync) {
868 		/* If I/O was synchronous, wait for it to complete. */
869 		rv = biowait(bp);
870 
871 		/* Release the buffer. */
872 		brelse(bp, 0);
873 
874 		return (rv);
875 	} else {
876 		return (0);
877 	}
878 }
879 
880 int
881 vn_bwrite(void *v)
882 {
883 	struct vop_bwrite_args *ap = v;
884 
885 	return (bwrite(ap->a_bp));
886 }
887 
888 /*
889  * Delayed write.
890  *
891  * The buffer is marked dirty, but is not queued for I/O.
892  * This routine should be used when the buffer is expected
893  * to be modified again soon, typically a small write that
894  * partially fills a buffer.
895  *
896  * NB: magnetic tapes cannot be delayed; they must be
897  * written in the order that the writes are requested.
898  *
899  * Described in Leffler, et al. (pp. 208-213).
900  */
901 void
902 bdwrite(buf_t *bp)
903 {
904 
905 	KASSERT(bp->b_vp == NULL || bp->b_vp->v_tag != VT_UFS ||
906 	    bp->b_vp->v_type == VBLK || ISSET(bp->b_flags, B_COWDONE));
907 	KASSERT(ISSET(bp->b_cflags, BC_BUSY));
908 	KASSERT(!cv_has_waiters(&bp->b_done));
909 
910 	/* If this is a tape block, write the block now. */
911 	if (bdev_type(bp->b_dev) == D_TAPE) {
912 		bawrite(bp);
913 		return;
914 	}
915 
916 	if (wapbl_vphaswapbl(bp->b_vp)) {
917 		struct mount *mp = wapbl_vptomp(bp->b_vp);
918 
919 		if (bp->b_iodone != mp->mnt_wapbl_op->wo_wapbl_biodone) {
920 			WAPBL_ADD_BUF(mp, bp);
921 		}
922 	}
923 
924 	/*
925 	 * If the block hasn't been seen before:
926 	 *	(1) Mark it as having been seen,
927 	 *	(2) Charge for the write,
928 	 *	(3) Make sure it's on its vnode's correct block list.
929 	 */
930 	KASSERT(bp->b_vp == NULL || bp->b_objlock == &bp->b_vp->v_interlock);
931 
932 	if (!ISSET(bp->b_oflags, BO_DELWRI)) {
933 		mutex_enter(&bufcache_lock);
934 		mutex_enter(bp->b_objlock);
935 		SET(bp->b_oflags, BO_DELWRI);
936 		curlwp->l_ru.ru_oublock++;
937 		reassignbuf(bp, bp->b_vp);
938 		mutex_exit(&bufcache_lock);
939 	} else {
940 		mutex_enter(bp->b_objlock);
941 	}
942 	/* Otherwise, the "write" is done, so mark and release the buffer. */
943 	CLR(bp->b_oflags, BO_DONE);
944 	mutex_exit(bp->b_objlock);
945 
946 	brelse(bp, 0);
947 }
948 
949 /*
950  * Asynchronous block write; just an asynchronous bwrite().
951  */
952 void
953 bawrite(buf_t *bp)
954 {
955 
956 	KASSERT(ISSET(bp->b_cflags, BC_BUSY));
957 
958 	SET(bp->b_flags, B_ASYNC);
959 	VOP_BWRITE(bp);
960 }
961 
962 /*
963  * Same as first half of bdwrite, mark buffer dirty, but do not release it.
964  * Call with the buffer interlock held.
965  *
966  * Note: called only from biodone() through ffs softdep's io_complete()
967  * Note2: smbfs also learned about bdirty().
968  */
969 void
970 bdirty(buf_t *bp)
971 {
972 
973 	KASSERT(mutex_owned(&bufcache_lock));
974 	KASSERT(bp->b_objlock == &bp->b_vp->v_interlock);
975 	KASSERT(mutex_owned(bp->b_objlock));
976 	KASSERT(ISSET(bp->b_cflags, BC_BUSY));
977 
978 	CLR(bp->b_cflags, BC_AGE);
979 
980 	if (!ISSET(bp->b_oflags, BO_DELWRI)) {
981 		SET(bp->b_oflags, BO_DELWRI);
982 		curlwp->l_ru.ru_oublock++;
983 		reassignbuf(bp, bp->b_vp);
984 	}
985 }
986 
987 
988 /*
989  * Release a buffer on to the free lists.
990  * Described in Bach (p. 46).
991  */
992 void
993 brelsel(buf_t *bp, int set)
994 {
995 	struct bqueue *bufq;
996 	struct vnode *vp;
997 
998 	KASSERT(mutex_owned(&bufcache_lock));
999 	KASSERT(!cv_has_waiters(&bp->b_done));
1000 	KASSERT(bp->b_refcnt > 0);
1001 
1002 	SET(bp->b_cflags, set);
1003 
1004 	KASSERT(ISSET(bp->b_cflags, BC_BUSY));
1005 	KASSERT(bp->b_iodone == NULL);
1006 
1007 	/* Wake up any processes waiting for any buffer to become free. */
1008 	cv_signal(&needbuffer_cv);
1009 
1010 	/* Wake up any proceeses waiting for _this_ buffer to become */
1011 	if (ISSET(bp->b_cflags, BC_WANTED))
1012 		CLR(bp->b_cflags, BC_WANTED|BC_AGE);
1013 
1014 	/*
1015 	 * Determine which queue the buffer should be on, then put it there.
1016 	 */
1017 
1018 	/* If it's locked, don't report an error; try again later. */
1019 	if (ISSET(bp->b_flags, B_LOCKED))
1020 		bp->b_error = 0;
1021 
1022 	/* If it's not cacheable, or an error, mark it invalid. */
1023 	if (ISSET(bp->b_cflags, BC_NOCACHE) || bp->b_error != 0)
1024 		SET(bp->b_cflags, BC_INVAL);
1025 
1026 	if (ISSET(bp->b_cflags, BC_VFLUSH)) {
1027 		/*
1028 		 * This is a delayed write buffer that was just flushed to
1029 		 * disk.  It is still on the LRU queue.  If it's become
1030 		 * invalid, then we need to move it to a different queue;
1031 		 * otherwise leave it in its current position.
1032 		 */
1033 		CLR(bp->b_cflags, BC_VFLUSH);
1034 		if (!ISSET(bp->b_cflags, BC_INVAL|BC_AGE) &&
1035 		    !ISSET(bp->b_flags, B_LOCKED) && bp->b_error == 0) {
1036 			KDASSERT(checkfreelist(bp, &bufqueues[BQ_LRU], 1));
1037 			goto already_queued;
1038 		} else {
1039 			bremfree(bp);
1040 		}
1041 	}
1042 
1043 	KDASSERT(checkfreelist(bp, &bufqueues[BQ_AGE], 0));
1044 	KDASSERT(checkfreelist(bp, &bufqueues[BQ_LRU], 0));
1045 	KDASSERT(checkfreelist(bp, &bufqueues[BQ_LOCKED], 0));
1046 
1047 	if ((bp->b_bufsize <= 0) || ISSET(bp->b_cflags, BC_INVAL)) {
1048 		/*
1049 		 * If it's invalid or empty, dissociate it from its vnode
1050 		 * and put on the head of the appropriate queue.
1051 		 */
1052 		if (bioopsp != NULL)
1053 			(*bioopsp->io_deallocate)(bp);
1054 
1055 		if (ISSET(bp->b_flags, B_LOCKED)) {
1056 			if (wapbl_vphaswapbl(vp = bp->b_vp)) {
1057 				struct mount *mp = wapbl_vptomp(vp);
1058 
1059 				KASSERT(bp->b_iodone
1060 				    != mp->mnt_wapbl_op->wo_wapbl_biodone);
1061 				WAPBL_REMOVE_BUF(mp, bp);
1062 			}
1063 		}
1064 
1065 		mutex_enter(bp->b_objlock);
1066 		CLR(bp->b_oflags, BO_DONE|BO_DELWRI);
1067 		if ((vp = bp->b_vp) != NULL) {
1068 			KASSERT(bp->b_objlock == &vp->v_interlock);
1069 			reassignbuf(bp, bp->b_vp);
1070 			brelvp(bp);
1071 			mutex_exit(&vp->v_interlock);
1072 		} else {
1073 			KASSERT(bp->b_objlock == &buffer_lock);
1074 			mutex_exit(bp->b_objlock);
1075 		}
1076 
1077 		if (bp->b_bufsize <= 0)
1078 			/* no data */
1079 			goto already_queued;
1080 		else
1081 			/* invalid data */
1082 			bufq = &bufqueues[BQ_AGE];
1083 		binsheadfree(bp, bufq);
1084 	} else  {
1085 		/*
1086 		 * It has valid data.  Put it on the end of the appropriate
1087 		 * queue, so that it'll stick around for as long as possible.
1088 		 * If buf is AGE, but has dependencies, must put it on last
1089 		 * bufqueue to be scanned, ie LRU. This protects against the
1090 		 * livelock where BQ_AGE only has buffers with dependencies,
1091 		 * and we thus never get to the dependent buffers in BQ_LRU.
1092 		 */
1093 		if (ISSET(bp->b_flags, B_LOCKED)) {
1094 			/* locked in core */
1095 			bufq = &bufqueues[BQ_LOCKED];
1096 		} else if (!ISSET(bp->b_cflags, BC_AGE)) {
1097 			/* valid data */
1098 			bufq = &bufqueues[BQ_LRU];
1099 		} else {
1100 			/* stale but valid data */
1101 			int has_deps;
1102 
1103 			if (bioopsp != NULL)
1104 				has_deps = (*bioopsp->io_countdeps)(bp, 0);
1105 			else
1106 				has_deps = 0;
1107 			bufq = has_deps ? &bufqueues[BQ_LRU] :
1108 			    &bufqueues[BQ_AGE];
1109 		}
1110 		binstailfree(bp, bufq);
1111 	}
1112 already_queued:
1113 	/* Unlock the buffer. */
1114 	CLR(bp->b_cflags, BC_AGE|BC_BUSY|BC_NOCACHE);
1115 	CLR(bp->b_flags, B_ASYNC);
1116 	cv_broadcast(&bp->b_busy);
1117 
1118 	if (bp->b_bufsize <= 0)
1119 		brele(bp);
1120 }
1121 
1122 void
1123 brelse(buf_t *bp, int set)
1124 {
1125 
1126 	mutex_enter(&bufcache_lock);
1127 	brelsel(bp, set);
1128 	mutex_exit(&bufcache_lock);
1129 }
1130 
1131 /*
1132  * Determine if a block is in the cache.
1133  * Just look on what would be its hash chain.  If it's there, return
1134  * a pointer to it, unless it's marked invalid.  If it's marked invalid,
1135  * we normally don't return the buffer, unless the caller explicitly
1136  * wants us to.
1137  */
1138 buf_t *
1139 incore(struct vnode *vp, daddr_t blkno)
1140 {
1141 	buf_t *bp;
1142 
1143 	KASSERT(mutex_owned(&bufcache_lock));
1144 
1145 	/* Search hash chain */
1146 	LIST_FOREACH(bp, BUFHASH(vp, blkno), b_hash) {
1147 		if (bp->b_lblkno == blkno && bp->b_vp == vp &&
1148 		    !ISSET(bp->b_cflags, BC_INVAL)) {
1149 		    	KASSERT(bp->b_objlock == &vp->v_interlock);
1150 		    	return (bp);
1151 		}
1152 	}
1153 
1154 	return (NULL);
1155 }
1156 
1157 /*
1158  * Get a block of requested size that is associated with
1159  * a given vnode and block offset. If it is found in the
1160  * block cache, mark it as having been found, make it busy
1161  * and return it. Otherwise, return an empty block of the
1162  * correct size. It is up to the caller to insure that the
1163  * cached blocks be of the correct size.
1164  */
1165 buf_t *
1166 getblk(struct vnode *vp, daddr_t blkno, int size, int slpflag, int slptimeo)
1167 {
1168 	int err, preserve;
1169 	buf_t *bp;
1170 
1171 	mutex_enter(&bufcache_lock);
1172  loop:
1173 	bp = incore(vp, blkno);
1174 	if (bp != NULL) {
1175 		err = bbusy(bp, ((slpflag & PCATCH) != 0), slptimeo, NULL);
1176 		if (err != 0) {
1177 			if (err == EPASSTHROUGH)
1178 				goto loop;
1179 			mutex_exit(&bufcache_lock);
1180 			return (NULL);
1181 		}
1182 		KASSERT(!cv_has_waiters(&bp->b_done));
1183 #ifdef DIAGNOSTIC
1184 		if (ISSET(bp->b_oflags, BO_DONE|BO_DELWRI) &&
1185 		    bp->b_bcount < size && vp->v_type != VBLK)
1186 			panic("getblk: block size invariant failed");
1187 #endif
1188 		bremfree(bp);
1189 		preserve = 1;
1190 	} else {
1191 		if ((bp = getnewbuf(slpflag, slptimeo, 0)) == NULL)
1192 			goto loop;
1193 
1194 		if (incore(vp, blkno) != NULL) {
1195 			/* The block has come into memory in the meantime. */
1196 			brelsel(bp, 0);
1197 			goto loop;
1198 		}
1199 
1200 		LIST_INSERT_HEAD(BUFHASH(vp, blkno), bp, b_hash);
1201 		bp->b_blkno = bp->b_lblkno = bp->b_rawblkno = blkno;
1202 		mutex_enter(&vp->v_interlock);
1203 		bgetvp(vp, bp);
1204 		mutex_exit(&vp->v_interlock);
1205 		preserve = 0;
1206 	}
1207 	mutex_exit(&bufcache_lock);
1208 
1209 	/*
1210 	 * LFS can't track total size of B_LOCKED buffer (locked_queue_bytes)
1211 	 * if we re-size buffers here.
1212 	 */
1213 	if (ISSET(bp->b_flags, B_LOCKED)) {
1214 		KASSERT(bp->b_bufsize >= size);
1215 	} else {
1216 		if (allocbuf(bp, size, preserve)) {
1217 			mutex_enter(&bufcache_lock);
1218 			LIST_REMOVE(bp, b_hash);
1219 			mutex_exit(&bufcache_lock);
1220 			brelse(bp, BC_INVAL);
1221 			return NULL;
1222 		}
1223 	}
1224 	BIO_SETPRIO(bp, BPRIO_DEFAULT);
1225 	return (bp);
1226 }
1227 
1228 /*
1229  * Get an empty, disassociated buffer of given size.
1230  */
1231 buf_t *
1232 geteblk(int size)
1233 {
1234 	buf_t *bp;
1235 	int error;
1236 
1237 	mutex_enter(&bufcache_lock);
1238 	while ((bp = getnewbuf(0, 0, 0)) == NULL)
1239 		;
1240 
1241 	SET(bp->b_cflags, BC_INVAL);
1242 	LIST_INSERT_HEAD(&invalhash, bp, b_hash);
1243 	mutex_exit(&bufcache_lock);
1244 	BIO_SETPRIO(bp, BPRIO_DEFAULT);
1245 	error = allocbuf(bp, size, 0);
1246 	KASSERT(error == 0);
1247 	return (bp);
1248 }
1249 
1250 /*
1251  * Expand or contract the actual memory allocated to a buffer.
1252  *
1253  * If the buffer shrinks, data is lost, so it's up to the
1254  * caller to have written it out *first*; this routine will not
1255  * start a write.  If the buffer grows, it's the callers
1256  * responsibility to fill out the buffer's additional contents.
1257  */
1258 int
1259 allocbuf(buf_t *bp, int size, int preserve)
1260 {
1261 	void *addr;
1262 	vsize_t oldsize, desired_size;
1263 	int oldcount;
1264 	int delta;
1265 
1266 	desired_size = buf_roundsize(size);
1267 	if (desired_size > MAXBSIZE)
1268 		printf("allocbuf: buffer larger than MAXBSIZE requested");
1269 
1270 	oldcount = bp->b_bcount;
1271 
1272 	bp->b_bcount = size;
1273 
1274 	oldsize = bp->b_bufsize;
1275 	if (oldsize == desired_size) {
1276 		/*
1277 		 * Do not short cut the WAPBL resize, as the buffer length
1278 		 * could still have changed and this would corrupt the
1279 		 * tracking of the transaction length.
1280 		 */
1281 		goto out;
1282 	}
1283 
1284 	/*
1285 	 * If we want a buffer of a different size, re-allocate the
1286 	 * buffer's memory; copy old content only if needed.
1287 	 */
1288 	addr = buf_malloc(desired_size);
1289 	if (addr == NULL)
1290 		return ENOMEM;
1291 	if (preserve)
1292 		memcpy(addr, bp->b_data, MIN(oldsize,desired_size));
1293 	if (bp->b_data != NULL)
1294 		buf_mrelease(bp->b_data, oldsize);
1295 	bp->b_data = addr;
1296 	bp->b_bufsize = desired_size;
1297 
1298 	/*
1299 	 * Update overall buffer memory counter (protected by bufcache_lock)
1300 	 */
1301 	delta = (long)desired_size - (long)oldsize;
1302 
1303 	mutex_enter(&bufcache_lock);
1304 	if ((bufmem += delta) > bufmem_hiwater) {
1305 		/*
1306 		 * Need to trim overall memory usage.
1307 		 */
1308 		while (buf_canrelease()) {
1309 			if (curcpu()->ci_schedstate.spc_flags &
1310 			    SPCF_SHOULDYIELD) {
1311 				mutex_exit(&bufcache_lock);
1312 				preempt();
1313 				mutex_enter(&bufcache_lock);
1314 			}
1315 			if (buf_trim() == 0)
1316 				break;
1317 		}
1318 	}
1319 	mutex_exit(&bufcache_lock);
1320 
1321  out:
1322 	if (wapbl_vphaswapbl(bp->b_vp))
1323 		WAPBL_RESIZE_BUF(wapbl_vptomp(bp->b_vp), bp, oldsize, oldcount);
1324 
1325 	return 0;
1326 }
1327 
1328 /*
1329  * Find a buffer which is available for use.
1330  * Select something from a free list.
1331  * Preference is to AGE list, then LRU list.
1332  *
1333  * Called with the buffer queues locked.
1334  * Return buffer locked.
1335  */
1336 buf_t *
1337 getnewbuf(int slpflag, int slptimeo, int from_bufq)
1338 {
1339 	buf_t *bp;
1340 	struct vnode *vp;
1341 
1342  start:
1343 	KASSERT(mutex_owned(&bufcache_lock));
1344 
1345 	/*
1346 	 * Get a new buffer from the pool.
1347 	 */
1348 	if (!from_bufq && buf_lotsfree()) {
1349 		mutex_exit(&bufcache_lock);
1350 		bp = pool_cache_get(buf_cache, PR_NOWAIT);
1351 		if (bp != NULL) {
1352 			memset((char *)bp, 0, sizeof(*bp));
1353 			buf_init(bp);
1354 			SET(bp->b_cflags, BC_BUSY);	/* mark buffer busy */
1355 			mutex_enter(&bufcache_lock);
1356 #if defined(DIAGNOSTIC)
1357 			bp->b_freelistindex = -1;
1358 #endif /* defined(DIAGNOSTIC) */
1359 			return (bp);
1360 		}
1361 		mutex_enter(&bufcache_lock);
1362 	}
1363 
1364 	KASSERT(mutex_owned(&bufcache_lock));
1365 	if ((bp = TAILQ_FIRST(&bufqueues[BQ_AGE].bq_queue)) != NULL ||
1366 	    (bp = TAILQ_FIRST(&bufqueues[BQ_LRU].bq_queue)) != NULL) {
1367 	    	KASSERT(!ISSET(bp->b_cflags, BC_BUSY) || ISSET(bp->b_cflags, BC_VFLUSH));
1368 		bremfree(bp);
1369 
1370 		/* Buffer is no longer on free lists. */
1371 		SET(bp->b_cflags, BC_BUSY);
1372 	} else {
1373 		/*
1374 		 * XXX: !from_bufq should be removed.
1375 		 */
1376 		if (!from_bufq || curlwp != uvm.pagedaemon_lwp) {
1377 			/* wait for a free buffer of any kind */
1378 			if ((slpflag & PCATCH) != 0)
1379 				(void)cv_timedwait_sig(&needbuffer_cv,
1380 				    &bufcache_lock, slptimeo);
1381 			else
1382 				(void)cv_timedwait(&needbuffer_cv,
1383 				    &bufcache_lock, slptimeo);
1384 		}
1385 		return (NULL);
1386 	}
1387 
1388 #ifdef DIAGNOSTIC
1389 	if (bp->b_bufsize <= 0)
1390 		panic("buffer %p: on queue but empty", bp);
1391 #endif
1392 
1393 	if (ISSET(bp->b_cflags, BC_VFLUSH)) {
1394 		/*
1395 		 * This is a delayed write buffer being flushed to disk.  Make
1396 		 * sure it gets aged out of the queue when it's finished, and
1397 		 * leave it off the LRU queue.
1398 		 */
1399 		CLR(bp->b_cflags, BC_VFLUSH);
1400 		SET(bp->b_cflags, BC_AGE);
1401 		goto start;
1402 	}
1403 
1404 	KASSERT(ISSET(bp->b_cflags, BC_BUSY));
1405 	KASSERT(bp->b_refcnt > 0);
1406     	KASSERT(!cv_has_waiters(&bp->b_done));
1407 
1408 	/*
1409 	 * If buffer was a delayed write, start it and return NULL
1410 	 * (since we might sleep while starting the write).
1411 	 */
1412 	if (ISSET(bp->b_oflags, BO_DELWRI)) {
1413 		/*
1414 		 * This buffer has gone through the LRU, so make sure it gets
1415 		 * reused ASAP.
1416 		 */
1417 		SET(bp->b_cflags, BC_AGE);
1418 		mutex_exit(&bufcache_lock);
1419 		bawrite(bp);
1420 		mutex_enter(&bufcache_lock);
1421 		return (NULL);
1422 	}
1423 
1424 	vp = bp->b_vp;
1425 	if (bioopsp != NULL)
1426 		(*bioopsp->io_deallocate)(bp);
1427 
1428 	/* clear out various other fields */
1429 	bp->b_cflags = BC_BUSY;
1430 	bp->b_oflags = 0;
1431 	bp->b_flags = 0;
1432 	bp->b_dev = NODEV;
1433 	bp->b_blkno = 0;
1434 	bp->b_lblkno = 0;
1435 	bp->b_rawblkno = 0;
1436 	bp->b_iodone = 0;
1437 	bp->b_error = 0;
1438 	bp->b_resid = 0;
1439 	bp->b_bcount = 0;
1440 
1441 	LIST_REMOVE(bp, b_hash);
1442 
1443 	/* Disassociate us from our vnode, if we had one... */
1444 	if (vp != NULL) {
1445 		mutex_enter(&vp->v_interlock);
1446 		brelvp(bp);
1447 		mutex_exit(&vp->v_interlock);
1448 	}
1449 
1450 	return (bp);
1451 }
1452 
1453 /*
1454  * Attempt to free an aged buffer off the queues.
1455  * Called with queue lock held.
1456  * Returns the amount of buffer memory freed.
1457  */
1458 static int
1459 buf_trim(void)
1460 {
1461 	buf_t *bp;
1462 	long size = 0;
1463 
1464 	KASSERT(mutex_owned(&bufcache_lock));
1465 
1466 	/* Instruct getnewbuf() to get buffers off the queues */
1467 	if ((bp = getnewbuf(PCATCH, 1, 1)) == NULL)
1468 		return 0;
1469 
1470 	KASSERT((bp->b_cflags & BC_WANTED) == 0);
1471 	size = bp->b_bufsize;
1472 	bufmem -= size;
1473 	if (size > 0) {
1474 		buf_mrelease(bp->b_data, size);
1475 		bp->b_bcount = bp->b_bufsize = 0;
1476 	}
1477 	/* brelse() will return the buffer to the global buffer pool */
1478 	brelsel(bp, 0);
1479 	return size;
1480 }
1481 
1482 int
1483 buf_drain(int n)
1484 {
1485 	int size = 0, sz;
1486 
1487 	KASSERT(mutex_owned(&bufcache_lock));
1488 
1489 	while (size < n && bufmem > bufmem_lowater) {
1490 		sz = buf_trim();
1491 		if (sz <= 0)
1492 			break;
1493 		size += sz;
1494 	}
1495 
1496 	return size;
1497 }
1498 
1499 /*
1500  * Wait for operations on the buffer to complete.
1501  * When they do, extract and return the I/O's error value.
1502  */
1503 int
1504 biowait(buf_t *bp)
1505 {
1506 
1507 	KASSERT(ISSET(bp->b_cflags, BC_BUSY));
1508 	KASSERT(bp->b_refcnt > 0);
1509 
1510 	mutex_enter(bp->b_objlock);
1511 	while (!ISSET(bp->b_oflags, BO_DONE | BO_DELWRI))
1512 		cv_wait(&bp->b_done, bp->b_objlock);
1513 	mutex_exit(bp->b_objlock);
1514 
1515 	return bp->b_error;
1516 }
1517 
1518 /*
1519  * Mark I/O complete on a buffer.
1520  *
1521  * If a callback has been requested, e.g. the pageout
1522  * daemon, do so. Otherwise, awaken waiting processes.
1523  *
1524  * [ Leffler, et al., says on p.247:
1525  *	"This routine wakes up the blocked process, frees the buffer
1526  *	for an asynchronous write, or, for a request by the pagedaemon
1527  *	process, invokes a procedure specified in the buffer structure" ]
1528  *
1529  * In real life, the pagedaemon (or other system processes) wants
1530  * to do async stuff to, and doesn't want the buffer brelse()'d.
1531  * (for swap pager, that puts swap buffers on the free lists (!!!),
1532  * for the vn device, that puts malloc'd buffers on the free lists!)
1533  */
1534 void
1535 biodone(buf_t *bp)
1536 {
1537 	int s;
1538 
1539 	KASSERT(!ISSET(bp->b_oflags, BO_DONE));
1540 
1541 	if (cpu_intr_p()) {
1542 		/* From interrupt mode: defer to a soft interrupt. */
1543 		s = splvm();
1544 		TAILQ_INSERT_TAIL(&curcpu()->ci_data.cpu_biodone, bp, b_actq);
1545 		softint_schedule(biodone_sih);
1546 		splx(s);
1547 	} else {
1548 		/* Process now - the buffer may be freed soon. */
1549 		biodone2(bp);
1550 	}
1551 }
1552 
1553 static void
1554 biodone2(buf_t *bp)
1555 {
1556 	void (*callout)(buf_t *);
1557 
1558 	if (bioopsp != NULL)
1559 		(*bioopsp->io_complete)(bp);
1560 
1561 	mutex_enter(bp->b_objlock);
1562 	/* Note that the transfer is done. */
1563 	if (ISSET(bp->b_oflags, BO_DONE))
1564 		panic("biodone2 already");
1565 	CLR(bp->b_flags, B_COWDONE);
1566 	SET(bp->b_oflags, BO_DONE);
1567 	BIO_SETPRIO(bp, BPRIO_DEFAULT);
1568 
1569 	/* Wake up waiting writers. */
1570 	if (!ISSET(bp->b_flags, B_READ))
1571 		vwakeup(bp);
1572 
1573 	if ((callout = bp->b_iodone) != NULL) {
1574 		/* Note callout done, then call out. */
1575 		KASSERT(!cv_has_waiters(&bp->b_done));
1576 		KERNEL_LOCK(1, NULL);		/* XXXSMP */
1577 		bp->b_iodone = NULL;
1578 		mutex_exit(bp->b_objlock);
1579 		(*callout)(bp);
1580 		KERNEL_UNLOCK_ONE(NULL);	/* XXXSMP */
1581 	} else if (ISSET(bp->b_flags, B_ASYNC)) {
1582 		/* If async, release. */
1583 		KASSERT(!cv_has_waiters(&bp->b_done));
1584 		mutex_exit(bp->b_objlock);
1585 		brelse(bp, 0);
1586 	} else {
1587 		/* Otherwise just wake up waiters in biowait(). */
1588 		cv_broadcast(&bp->b_done);
1589 		mutex_exit(bp->b_objlock);
1590 	}
1591 }
1592 
1593 static void
1594 biointr(void *cookie)
1595 {
1596 	struct cpu_info *ci;
1597 	buf_t *bp;
1598 	int s;
1599 
1600 	ci = curcpu();
1601 
1602 	while (!TAILQ_EMPTY(&ci->ci_data.cpu_biodone)) {
1603 		KASSERT(curcpu() == ci);
1604 
1605 		s = splvm();
1606 		bp = TAILQ_FIRST(&ci->ci_data.cpu_biodone);
1607 		TAILQ_REMOVE(&ci->ci_data.cpu_biodone, bp, b_actq);
1608 		splx(s);
1609 
1610 		biodone2(bp);
1611 	}
1612 }
1613 
1614 /*
1615  * Return a count of buffers on the "locked" queue.
1616  */
1617 int
1618 count_lock_queue(void)
1619 {
1620 	buf_t *bp;
1621 	int n = 0;
1622 
1623 	mutex_enter(&bufcache_lock);
1624 	TAILQ_FOREACH(bp, &bufqueues[BQ_LOCKED].bq_queue, b_freelist)
1625 		n++;
1626 	mutex_exit(&bufcache_lock);
1627 	return (n);
1628 }
1629 
1630 /*
1631  * Wait for all buffers to complete I/O
1632  * Return the number of "stuck" buffers.
1633  */
1634 int
1635 buf_syncwait(void)
1636 {
1637 	buf_t *bp;
1638 	int iter, nbusy, nbusy_prev = 0, dcount, ihash;
1639 
1640 	dcount = 10000;
1641 	for (iter = 0; iter < 20;) {
1642 		mutex_enter(&bufcache_lock);
1643 		nbusy = 0;
1644 		for (ihash = 0; ihash < bufhash+1; ihash++) {
1645 		    LIST_FOREACH(bp, &bufhashtbl[ihash], b_hash) {
1646 			if ((bp->b_cflags & (BC_BUSY|BC_INVAL)) == BC_BUSY)
1647 				nbusy += ((bp->b_flags & B_READ) == 0);
1648 			/*
1649 			 * With soft updates, some buffers that are
1650 			 * written will be remarked as dirty until other
1651 			 * buffers are written.
1652 			 */
1653 			if (bp->b_vp && bp->b_vp->v_tag != VT_NON
1654 			    && bp->b_vp->v_mount
1655 			    && (bp->b_vp->v_mount->mnt_flag & MNT_SOFTDEP)
1656 			    && (bp->b_oflags & BO_DELWRI)) {
1657 				bremfree(bp);
1658 				bp->b_cflags |= BC_BUSY;
1659 				nbusy++;
1660 				mutex_exit(&bufcache_lock);
1661 				bawrite(bp);
1662 				if (dcount-- <= 0) {
1663 					printf("softdep ");
1664 					goto fail;
1665 				}
1666 				mutex_enter(&bufcache_lock);
1667 			}
1668 		    }
1669 		}
1670 		mutex_exit(&bufcache_lock);
1671 
1672 		if (nbusy == 0)
1673 			break;
1674 		if (nbusy_prev == 0)
1675 			nbusy_prev = nbusy;
1676 		printf("%d ", nbusy);
1677 		kpause("bflush", false, (iter == 0) ? 1 : hz / 25 * iter, NULL);
1678 		if (nbusy >= nbusy_prev) /* we didn't flush anything */
1679 			iter++;
1680 		else
1681 			nbusy_prev = nbusy;
1682 	}
1683 
1684 	if (nbusy) {
1685 fail:;
1686 #if defined(DEBUG) || defined(DEBUG_HALT_BUSY)
1687 		printf("giving up\nPrinting vnodes for busy buffers\n");
1688 		for (ihash = 0; ihash < bufhash+1; ihash++) {
1689 		    LIST_FOREACH(bp, &bufhashtbl[ihash], b_hash) {
1690 			if ((bp->b_cflags & (BC_BUSY|BC_INVAL)) == BC_BUSY &&
1691 			    (bp->b_flags & B_READ) == 0)
1692 				vprint(NULL, bp->b_vp);
1693 		    }
1694 		}
1695 #endif
1696 	}
1697 
1698 	return nbusy;
1699 }
1700 
1701 static void
1702 sysctl_fillbuf(buf_t *i, struct buf_sysctl *o)
1703 {
1704 
1705 	o->b_flags = i->b_flags | i->b_cflags | i->b_oflags;
1706 	o->b_error = i->b_error;
1707 	o->b_prio = i->b_prio;
1708 	o->b_dev = i->b_dev;
1709 	o->b_bufsize = i->b_bufsize;
1710 	o->b_bcount = i->b_bcount;
1711 	o->b_resid = i->b_resid;
1712 	o->b_addr = PTRTOUINT64(i->b_data);
1713 	o->b_blkno = i->b_blkno;
1714 	o->b_rawblkno = i->b_rawblkno;
1715 	o->b_iodone = PTRTOUINT64(i->b_iodone);
1716 	o->b_proc = PTRTOUINT64(i->b_proc);
1717 	o->b_vp = PTRTOUINT64(i->b_vp);
1718 	o->b_saveaddr = PTRTOUINT64(i->b_saveaddr);
1719 	o->b_lblkno = i->b_lblkno;
1720 }
1721 
1722 #define KERN_BUFSLOP 20
1723 static int
1724 sysctl_dobuf(SYSCTLFN_ARGS)
1725 {
1726 	buf_t *bp;
1727 	struct buf_sysctl bs;
1728 	struct bqueue *bq;
1729 	char *dp;
1730 	u_int i, op, arg;
1731 	size_t len, needed, elem_size, out_size;
1732 	int error, elem_count, retries;
1733 
1734 	if (namelen == 1 && name[0] == CTL_QUERY)
1735 		return (sysctl_query(SYSCTLFN_CALL(rnode)));
1736 
1737 	if (namelen != 4)
1738 		return (EINVAL);
1739 
1740 	retries = 100;
1741  retry:
1742 	dp = oldp;
1743 	len = (oldp != NULL) ? *oldlenp : 0;
1744 	op = name[0];
1745 	arg = name[1];
1746 	elem_size = name[2];
1747 	elem_count = name[3];
1748 	out_size = MIN(sizeof(bs), elem_size);
1749 
1750 	/*
1751 	 * at the moment, these are just "placeholders" to make the
1752 	 * API for retrieving kern.buf data more extensible in the
1753 	 * future.
1754 	 *
1755 	 * XXX kern.buf currently has "netbsd32" issues.  hopefully
1756 	 * these will be resolved at a later point.
1757 	 */
1758 	if (op != KERN_BUF_ALL || arg != KERN_BUF_ALL ||
1759 	    elem_size < 1 || elem_count < 0)
1760 		return (EINVAL);
1761 
1762 	error = 0;
1763 	needed = 0;
1764 	sysctl_unlock();
1765 	mutex_enter(&bufcache_lock);
1766 	for (i = 0; i < BQUEUES; i++) {
1767 		bq = &bufqueues[i];
1768 		TAILQ_FOREACH(bp, &bq->bq_queue, b_freelist) {
1769 			bq->bq_marker = bp;
1770 			if (len >= elem_size && elem_count > 0) {
1771 				sysctl_fillbuf(bp, &bs);
1772 				mutex_exit(&bufcache_lock);
1773 				error = copyout(&bs, dp, out_size);
1774 				mutex_enter(&bufcache_lock);
1775 				if (error)
1776 					break;
1777 				if (bq->bq_marker != bp) {
1778 					/*
1779 					 * This sysctl node is only for
1780 					 * statistics.  Retry; if the
1781 					 * queue keeps changing, then
1782 					 * bail out.
1783 					 */
1784 					if (retries-- == 0) {
1785 						error = EAGAIN;
1786 						break;
1787 					}
1788 					mutex_exit(&bufcache_lock);
1789 					goto retry;
1790 				}
1791 				dp += elem_size;
1792 				len -= elem_size;
1793 			}
1794 			if (elem_count > 0) {
1795 				needed += elem_size;
1796 				if (elem_count != INT_MAX)
1797 					elem_count--;
1798 			}
1799 		}
1800 		if (error != 0)
1801 			break;
1802 	}
1803 	mutex_exit(&bufcache_lock);
1804 	sysctl_relock();
1805 
1806 	*oldlenp = needed;
1807 	if (oldp == NULL)
1808 		*oldlenp += KERN_BUFSLOP * sizeof(buf_t);
1809 
1810 	return (error);
1811 }
1812 
1813 static int
1814 sysctl_bufvm_update(SYSCTLFN_ARGS)
1815 {
1816 	int t, error, rv;
1817 	struct sysctlnode node;
1818 
1819 	node = *rnode;
1820 	node.sysctl_data = &t;
1821 	t = *(int *)rnode->sysctl_data;
1822 	error = sysctl_lookup(SYSCTLFN_CALL(&node));
1823 	if (error || newp == NULL)
1824 		return (error);
1825 
1826 	if (t < 0)
1827 		return EINVAL;
1828 	if (rnode->sysctl_data == &bufcache) {
1829 		if (t > 100)
1830 			return (EINVAL);
1831 		bufcache = t;
1832 		buf_setwm();
1833 	} else if (rnode->sysctl_data == &bufmem_lowater) {
1834 		if (bufmem_hiwater - t < 16)
1835 			return (EINVAL);
1836 		bufmem_lowater = t;
1837 	} else if (rnode->sysctl_data == &bufmem_hiwater) {
1838 		if (t - bufmem_lowater < 16)
1839 			return (EINVAL);
1840 		bufmem_hiwater = t;
1841 	} else
1842 		return (EINVAL);
1843 
1844 	/* Drain until below new high water mark */
1845 	sysctl_unlock();
1846 	mutex_enter(&bufcache_lock);
1847 	while ((t = bufmem - bufmem_hiwater) >= 0) {
1848 		rv = buf_drain(t / (2 * 1024));
1849 		if (rv <= 0)
1850 			break;
1851 	}
1852 	mutex_exit(&bufcache_lock);
1853 	sysctl_relock();
1854 
1855 	return 0;
1856 }
1857 
1858 static struct sysctllog *vfsbio_sysctllog;
1859 
1860 static void
1861 sysctl_kern_buf_setup(void)
1862 {
1863 
1864 	sysctl_createv(&vfsbio_sysctllog, 0, NULL, NULL,
1865 		       CTLFLAG_PERMANENT,
1866 		       CTLTYPE_NODE, "kern", NULL,
1867 		       NULL, 0, NULL, 0,
1868 		       CTL_KERN, CTL_EOL);
1869 	sysctl_createv(&vfsbio_sysctllog, 0, NULL, NULL,
1870 		       CTLFLAG_PERMANENT,
1871 		       CTLTYPE_NODE, "buf",
1872 		       SYSCTL_DESCR("Kernel buffer cache information"),
1873 		       sysctl_dobuf, 0, NULL, 0,
1874 		       CTL_KERN, KERN_BUF, CTL_EOL);
1875 }
1876 
1877 static void
1878 sysctl_vm_buf_setup(void)
1879 {
1880 
1881 	sysctl_createv(&vfsbio_sysctllog, 0, NULL, NULL,
1882 		       CTLFLAG_PERMANENT,
1883 		       CTLTYPE_NODE, "vm", NULL,
1884 		       NULL, 0, NULL, 0,
1885 		       CTL_VM, CTL_EOL);
1886 	sysctl_createv(&vfsbio_sysctllog, 0, NULL, NULL,
1887 		       CTLFLAG_PERMANENT|CTLFLAG_READWRITE,
1888 		       CTLTYPE_INT, "bufcache",
1889 		       SYSCTL_DESCR("Percentage of physical memory to use for "
1890 				    "buffer cache"),
1891 		       sysctl_bufvm_update, 0, &bufcache, 0,
1892 		       CTL_VM, CTL_CREATE, CTL_EOL);
1893 	sysctl_createv(&vfsbio_sysctllog, 0, NULL, NULL,
1894 		       CTLFLAG_PERMANENT|CTLFLAG_READONLY,
1895 		       CTLTYPE_INT, "bufmem",
1896 		       SYSCTL_DESCR("Amount of kernel memory used by buffer "
1897 				    "cache"),
1898 		       NULL, 0, &bufmem, 0,
1899 		       CTL_VM, CTL_CREATE, CTL_EOL);
1900 	sysctl_createv(&vfsbio_sysctllog, 0, NULL, NULL,
1901 		       CTLFLAG_PERMANENT|CTLFLAG_READWRITE,
1902 		       CTLTYPE_INT, "bufmem_lowater",
1903 		       SYSCTL_DESCR("Minimum amount of kernel memory to "
1904 				    "reserve for buffer cache"),
1905 		       sysctl_bufvm_update, 0, &bufmem_lowater, 0,
1906 		       CTL_VM, CTL_CREATE, CTL_EOL);
1907 	sysctl_createv(&vfsbio_sysctllog, 0, NULL, NULL,
1908 		       CTLFLAG_PERMANENT|CTLFLAG_READWRITE,
1909 		       CTLTYPE_INT, "bufmem_hiwater",
1910 		       SYSCTL_DESCR("Maximum amount of kernel memory to use "
1911 				    "for buffer cache"),
1912 		       sysctl_bufvm_update, 0, &bufmem_hiwater, 0,
1913 		       CTL_VM, CTL_CREATE, CTL_EOL);
1914 }
1915 
1916 #ifdef DEBUG
1917 /*
1918  * Print out statistics on the current allocation of the buffer pool.
1919  * Can be enabled to print out on every ``sync'' by setting "syncprt"
1920  * in vfs_syscalls.c using sysctl.
1921  */
1922 void
1923 vfs_bufstats(void)
1924 {
1925 	int i, j, count;
1926 	buf_t *bp;
1927 	struct bqueue *dp;
1928 	int counts[(MAXBSIZE / PAGE_SIZE) + 1];
1929 	static const char *bname[BQUEUES] = { "LOCKED", "LRU", "AGE" };
1930 
1931 	for (dp = bufqueues, i = 0; dp < &bufqueues[BQUEUES]; dp++, i++) {
1932 		count = 0;
1933 		for (j = 0; j <= MAXBSIZE/PAGE_SIZE; j++)
1934 			counts[j] = 0;
1935 		TAILQ_FOREACH(bp, &dp->bq_queue, b_freelist) {
1936 			counts[bp->b_bufsize/PAGE_SIZE]++;
1937 			count++;
1938 		}
1939 		printf("%s: total-%d", bname[i], count);
1940 		for (j = 0; j <= MAXBSIZE/PAGE_SIZE; j++)
1941 			if (counts[j] != 0)
1942 				printf(", %d-%d", j * PAGE_SIZE, counts[j]);
1943 		printf("\n");
1944 	}
1945 }
1946 #endif /* DEBUG */
1947 
1948 /* ------------------------------ */
1949 
1950 buf_t *
1951 getiobuf(struct vnode *vp, bool waitok)
1952 {
1953 	buf_t *bp;
1954 
1955 	bp = pool_cache_get(bufio_cache, (waitok ? PR_WAITOK : PR_NOWAIT));
1956 	if (bp == NULL)
1957 		return bp;
1958 
1959 	buf_init(bp);
1960 
1961 	if ((bp->b_vp = vp) == NULL)
1962 		bp->b_objlock = &buffer_lock;
1963 	else
1964 		bp->b_objlock = &vp->v_interlock;
1965 
1966 	return bp;
1967 }
1968 
1969 void
1970 putiobuf(buf_t *bp)
1971 {
1972 
1973 	buf_destroy(bp);
1974 	pool_cache_put(bufio_cache, bp);
1975 }
1976 
1977 /*
1978  * nestiobuf_iodone: b_iodone callback for nested buffers.
1979  */
1980 
1981 void
1982 nestiobuf_iodone(buf_t *bp)
1983 {
1984 	buf_t *mbp = bp->b_private;
1985 	int error;
1986 	int donebytes;
1987 
1988 	KASSERT(bp->b_bcount <= bp->b_bufsize);
1989 	KASSERT(mbp != bp);
1990 
1991 	error = bp->b_error;
1992 	if (bp->b_error == 0 &&
1993 	    (bp->b_bcount < bp->b_bufsize || bp->b_resid > 0)) {
1994 		/*
1995 		 * Not all got transfered, raise an error. We have no way to
1996 		 * propagate these conditions to mbp.
1997 		 */
1998 		error = EIO;
1999 	}
2000 
2001 	donebytes = bp->b_bufsize;
2002 
2003 	putiobuf(bp);
2004 	nestiobuf_done(mbp, donebytes, error);
2005 }
2006 
2007 /*
2008  * nestiobuf_setup: setup a "nested" buffer.
2009  *
2010  * => 'mbp' is a "master" buffer which is being divided into sub pieces.
2011  * => 'bp' should be a buffer allocated by getiobuf.
2012  * => 'offset' is a byte offset in the master buffer.
2013  * => 'size' is a size in bytes of this nested buffer.
2014  */
2015 
2016 void
2017 nestiobuf_setup(buf_t *mbp, buf_t *bp, int offset, size_t size)
2018 {
2019 	const int b_read = mbp->b_flags & B_READ;
2020 	struct vnode *vp = mbp->b_vp;
2021 
2022 	KASSERT(mbp->b_bcount >= offset + size);
2023 	bp->b_vp = vp;
2024 	bp->b_dev = mbp->b_dev;
2025 	bp->b_objlock = mbp->b_objlock;
2026 	bp->b_cflags = BC_BUSY;
2027 	bp->b_flags = B_ASYNC | b_read;
2028 	bp->b_iodone = nestiobuf_iodone;
2029 	bp->b_data = (char *)mbp->b_data + offset;
2030 	bp->b_resid = bp->b_bcount = size;
2031 	bp->b_bufsize = bp->b_bcount;
2032 	bp->b_private = mbp;
2033 	BIO_COPYPRIO(bp, mbp);
2034 	if (!b_read && vp != NULL) {
2035 		mutex_enter(&vp->v_interlock);
2036 		vp->v_numoutput++;
2037 		mutex_exit(&vp->v_interlock);
2038 	}
2039 }
2040 
2041 /*
2042  * nestiobuf_done: propagate completion to the master buffer.
2043  *
2044  * => 'donebytes' specifies how many bytes in the 'mbp' is completed.
2045  * => 'error' is an errno(2) that 'donebytes' has been completed with.
2046  */
2047 
2048 void
2049 nestiobuf_done(buf_t *mbp, int donebytes, int error)
2050 {
2051 
2052 	if (donebytes == 0) {
2053 		return;
2054 	}
2055 	mutex_enter(mbp->b_objlock);
2056 	KASSERT(mbp->b_resid >= donebytes);
2057 	mbp->b_resid -= donebytes;
2058 	if (error)
2059 		mbp->b_error = error;
2060 	if (mbp->b_resid == 0) {
2061 		mutex_exit(mbp->b_objlock);
2062 		biodone(mbp);
2063 	} else
2064 		mutex_exit(mbp->b_objlock);
2065 }
2066 
2067 void
2068 buf_init(buf_t *bp)
2069 {
2070 
2071 	LIST_INIT(&bp->b_dep);
2072 	cv_init(&bp->b_busy, "biolock");
2073 	cv_init(&bp->b_done, "biowait");
2074 	bp->b_dev = NODEV;
2075 	bp->b_error = 0;
2076 	bp->b_flags = 0;
2077 	bp->b_cflags = 0;
2078 	bp->b_oflags = 0;
2079 	bp->b_objlock = &buffer_lock;
2080 	bp->b_iodone = NULL;
2081 	bp->b_refcnt = 1;
2082 	bp->b_dev = NODEV;
2083 	bp->b_vnbufs.le_next = NOLIST;
2084 	BIO_SETPRIO(bp, BPRIO_DEFAULT);
2085 }
2086 
2087 void
2088 buf_destroy(buf_t *bp)
2089 {
2090 
2091 	cv_destroy(&bp->b_done);
2092 	cv_destroy(&bp->b_busy);
2093 }
2094 
2095 int
2096 bbusy(buf_t *bp, bool intr, int timo, kmutex_t *interlock)
2097 {
2098 	int error;
2099 
2100 	KASSERT(mutex_owned(&bufcache_lock));
2101 
2102 	if ((bp->b_cflags & BC_BUSY) != 0) {
2103 		if (curlwp == uvm.pagedaemon_lwp)
2104 			return EDEADLK;
2105 		bp->b_cflags |= BC_WANTED;
2106 		bref(bp);
2107 		if (interlock != NULL)
2108 			mutex_exit(interlock);
2109 		if (intr) {
2110 			error = cv_timedwait_sig(&bp->b_busy, &bufcache_lock,
2111 			    timo);
2112 		} else {
2113 			error = cv_timedwait(&bp->b_busy, &bufcache_lock,
2114 			    timo);
2115 		}
2116 		brele(bp);
2117 		if (interlock != NULL)
2118 			mutex_enter(interlock);
2119 		if (error != 0)
2120 			return error;
2121 		return EPASSTHROUGH;
2122 	}
2123 	bp->b_cflags |= BC_BUSY;
2124 
2125 	return 0;
2126 }
2127